Improve chevrotain parser error recovery

[msujew]

While looking into the Chevrotain documentation on fault tolerance, I've noticed that we're not supposed to catch the errors thrown by Chevrotain ourselves, but let them propagate. We've been effectively starving the catch block in this method, which is in charge of actually recovering from the error.

This results in some less-than-optimal behavior from Langium when it comes to error recovery. I've been playing around with this change using our example languages and it seems to behave pretty well. Refer to the additional comments to see why some of the changes were required.

[msujew]

@cdietrich Are you interested in testing this? I believe this should be beneficial for most adopters of Langium. I can create a next release of this change if you want to.

[msujew]

This PR is available as 3.4.0-next.10c0eb7 on npm.

[spoenemann]

I haven't seen any difference in error recovery with the example languages. Do you have an example where the recovery behaves differently?

One minor difference: in the statemachine example, when I insert state state between the events and the initialState, the previous version showed two parser errors

Expecting token of type 'initialState' but found state.statemachine
Expecting end of file but found state.

while the new version shows only one. But in both versions, the recovery does not work in this case: I get no cross-reference resolution for the rest of the file.

[msujew]

I haven't seen any difference in error recovery with the example languages. Do you have an example where the recovery behaves differently?

Me neither (at least with the examples). It might be that the only difference actually is related to the improved parser error message. Thinking further about this change, it's likely that this change mainly improves the error message behavior, as we likely end up at the same recovered parser state, but with less steps (and therefore fewer parser errors).